{"id":4655,"date":"2021-06-17T10:27:40","date_gmt":"2021-06-17T14:27:40","guid":{"rendered":"https:\/\/blogs.mathworks.com\/steve\/?p=4655"},"modified":"2021-06-17T10:28:24","modified_gmt":"2021-06-17T14:28:24","slug":"use-imbinarize-to-threshold-gray-scale-images","status":"publish","type":"post","link":"https:\/\/blogs.mathworks.com\/steve\/2021\/06\/17\/use-imbinarize-to-threshold-gray-scale-images\/","title":{"rendered":"Use imbinarize to Threshold Gray-Scale Images"},"content":{"rendered":"
Today, I want to convince you to use <\/span>imbinarize<\/span> instead of <\/span>im2bw<\/span>.<\/span><\/div>
Background: I recently saw some data suggesting that many Image Processing Toolbox users are still using <\/span>im2bw<\/span>, an old function that dates back to the original toolbox release in 1993. We recommend using the newer function, <\/span>imbinarize<\/span><\/a>, because it saves a step in the most common scenario, and because it offers additional flexibility if you need it.<\/span><\/div>
I'll explain by updating an old blog post topic called <\/span>\"Tracing George,\"<\/span><\/a> in which I try to extract the curve shown in the following image:<\/span><\/div>
url = <\/span>'https:\/\/blogs.mathworks.com\/images\/steve\/36\/george.jpg'<\/span>;<\/span><\/span><\/div><\/div>
I = imread(url);<\/span><\/span><\/div><\/div>
imshow(I)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
I call this \"George.\" I created George originally for some examples and diagrams in the book <\/span>Digital Image Processing Using MATLAB<\/span>. I found at some point that I needed to reproduce the outline curve, but I didn't have the original curve data. (Silly me.) So, I experimented with ways to extract the original curve data from the image.<\/span><\/div>
An obvious first step is to threshold the image. In my original blog post, I used two functions together:<\/span><\/div>
threshold = graythresh(I);<\/span><\/span><\/div>
bw = im2bw(I,threshold)<\/span><\/span><\/div><\/div>
The function <\/span>graythresh<\/span> is used to compute an \"optimal\" threshold value (optimal according to certain criteria). The function <\/span>im2bw<\/span> can be used without specifying the second argument, but then it just uses a fixed threshold value that usually isn't satisfactory:<\/span><\/div>
bw = im2bw(I);<\/span><\/span><\/div><\/div>
imshow(bw)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
The result with <\/span>graythresh<\/span> is much better.<\/span><\/div>
t = graythresh(I)<\/span><\/span><\/div>
t = 0.4392<\/div><\/div><\/div>
bw2 = im2bw(I,t);<\/span><\/span><\/div><\/div>
imshow(bw2)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
Over time, we observed that almost everyone (including us) always used <\/span>im2bw<\/span> and <\/span>graythresh<\/span> together. That caused to reconsider the functional design. The replacement function, <\/span>imbinarize<\/span>, uses <\/span>graythresh<\/span> automatically.<\/span><\/div>
bw3 = imbinarize(I);<\/span><\/span><\/div><\/div>
imshow(bw3)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
As a side note, here's the rest of the code that I used to extract the curve. See my original <\/span>\"Tracing George\" blog post <\/span><\/a>for an explanation.<\/span><\/div>
bw4 = ~bw3;<\/span><\/span><\/div><\/div>
bw5 = bwmorph(bw4, <\/span>'thin'<\/span>, inf);<\/span><\/span><\/div><\/div>
boundaries = bwboundaries(bw5);<\/span><\/span><\/div><\/div>
b = boundaries{1};<\/span><\/span><\/div><\/div>
b = b(1:floor(end\/2), :);<\/span><\/span><\/div><\/div>
x = b(:,2);<\/span><\/span><\/div><\/div>
y = b(:,1);<\/span><\/span><\/div><\/div>
<\/div><\/div>
plot(x,y)<\/span><\/span><\/div><\/div>
axis <\/span>ij<\/span><\/span><\/div><\/div>
axis <\/span>equal<\/span><\/span><\/div><\/div>
title(<\/span>'George P. Burdell'<\/span>)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>

Locally Adaptive Thresholding<\/span><\/h2>
For more difficult thresholding challenges, <\/span>imbinarize<\/span> also supports a locally adaptive method. To demonstrate, here's a picture of a page from Digital Image Processing Using MATLAB that I took in my office just now, with uneven lighting.<\/span><\/div>
page_url = <\/span>\"https:\/\/blogs.mathworks.com\/steve\/files\/dipum3e-page-388-gray.jpg\"<\/span>;<\/span><\/span><\/div><\/div>
A = imread(page_url);<\/span><\/span><\/div><\/div>
imshow(A)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
Because of the uneven lighting, even the best global threshold doesn't produce a suitable result.<\/span><\/div>
page_bw = imbinarize(A);<\/span><\/span><\/div><\/div>
imshow(page_bw)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
You can see that there is an area of missing text at the upper right. A locally adaptive thresholding method would be better. The function <\/span>imbinarize<\/span> supports a locally adaptive method, which is specified using the <\/span>\"adaptive\"<\/span> input option. With locally adaptive methods, though, it is very helpful to specify whether the foreground lighter or darker than the background. You can tell <\/span>imbinarize<\/span> that information using the <\/span>\"ForegroundPolarity\"<\/span> option.<\/span><\/div>
Here it is in action.<\/span><\/div>
page_bw2 = imbinarize(A,<\/span>\"adaptive\"<\/span>,<\/span>\"ForegroundPolarity\"<\/span>,<\/span>\"dark\"<\/span>);<\/span><\/span><\/div><\/div>
imshow(page_bw2)<\/span><\/span><\/div>
<\/div><\/div><\/div><\/div><\/div>
You can see that now all of the text has been cleanly included in the resulting binary image.<\/span><\/div>
So, the next time you go reaching for <\/span>im2bw<\/span> and <\/span>graythresh<\/span>, try just using <\/span>imbinarize<\/span> instead.<\/span><\/div>\r\n<\/div>